[Relax][ONNX] Support ConcatFromSequence new_axis=1 and refactor sequ…

[Dayuxiaoshui]

Summary

• Implement ONNX ConcatFromSequence with new_axis=1: expand_dims along axis, then concat on the same axis (matches ONNX Runtime).
• Add _onnx_sequence_tuple_as_tensor_list for SequenceErase, SequenceInsert, and ConcatFromSequence.
• SequenceInsert: coerce position to int and validate insert index.

Testing

• Parametrize new_axis / axis (including axis=-1); fix expected shape for axis=1.
• Add three-tensor new_axis=1 case; re-run SequenceInsert / SequenceErase.
  [Tracking Issue][ONNX] Complete missing and limited operators in ONNX frontend #18945

[Dayuxiaoshui]

cc @tlopex

[gemini-code-assist]

Code Review

This pull request introduces a helper function to standardize ONNX sequence handling and implements support for the new_axis attribute in ConcatFromSequence. It also adds bounds checking for SequenceInsert and expands test coverage. Review feedback suggests using more idiomatic Python for tuple conversion and improving efficiency in SequenceErase by using in-place deletion.

[tlopex]

Thanks for the PR. A few comments before merge:

1. _onnx_sequence_tuple_as_tensor_list feels unnecessary since it only does list(sequence). I’d just inline it.

2. For the insert bounds check, please add a short comment that position == seq_len is valid since insert can append to the end.

3. In SequenceErase, del items[position] is okay, but I’d prefer an assert position >= 0 for safety, or just keep the original list-comprehension version.

4. Maybe also add a basic axis=0, new_axis=0 case.

[Dayuxiaoshui]

cc @tlopex